IBM Granite Embedding Small English R2
IBM · USA · 2025
The most downloaded model IBM has ever published, and it cannot hold a conversation: 47 million parameters that turn text into search vectors, pulled five times more often than the company's flagship chat model.
Granite-Embedding-Small-English-R2 is not a chatbot and cannot answer a question. It reads a piece of text and returns a list of 384 numbers — a position in space where similar meanings land close together. That is the machinery behind site search, document retrieval and every system that feeds a language model the right paragraph before it answers. It is also, by a wide margin, the most used model IBM has ever released. The repository was pulled about 5.75 million times in the thirty days to 1 September 2026 — roughly five times more than Granite 4.1 8B, the company's most popular chat model, and more than every Granite chat model in the catalogue put together. The reason is mundane: a retrieval system calls its embedding model on every document it indexes and every query it receives, while a chat model is called once per conversation. The model is small on purpose. 47 million parameters, twelve layers, an output vector of 384 numbers and a context window of 8,192 tokens — IBM calls it a first of its kind at that size, and it encodes about 199 documents per second on a single H100. Its larger sibling, granite-embedding-english-r2, has 149 million parameters and 768-number vectors, scores two to three points higher on most retrieval benchmarks and is a third slower. The architecture is ModernBERT: alternating attention windows, rotary positions, GeGLU activations, no bias terms, a 50,368-token vocabulary trained on code and prose together. On IBM's own runs it reaches 50.9 on BEIR retrieval, 61.1 on MTEB-v2, 53.8 on code retrieval and 39.8 on long-document search — close enough to the 125-million-parameter previous generation that it replaces, at a third of the size. The part worth reading twice is what is missing from the training data. IBM states outright that it did not use MS MARCO, the retrieval dataset most open embedding models are trained on, because its licence forbids commercial use. That decision costs accuracy and IBM took it anyway — the same trade-off the company documented in August 2026, when it published a speech model twice over and forbade commercial use of the better version. The licence is plain Apache 2.0 with no territorial carve-out. English only: longer or non-English text is truncated or handled poorly, and Polish users need the multilingual R2 model instead.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!